Microsoft researchers have released a new AI framework called Auto Evol-Instruct that can automatically evolve guidance data sets without any human intervention. This is of great significance for improving the ability of large language models (LLMs) to follow complex instructions. Traditional evolutionary methods rely on artificially designed rules, which are inefficient and difficult to adapt to new tasks. Auto Evol-Instruct, on the other hand, automatically analyzes instructions through LLMs, independently designs and optimizes evolution rules, realizes an automated and efficient evolution process, and greatly improves the complexity and diversity of data sets.
Recently, Microsoft researchers proposed a new AI framework called Auto Evol-Instruct, which can automatically evolve guidance data sets without any human intervention.
In the field of artificial intelligence, the development of large language models (LLMs) is crucial, especially in improving the ability of these models to follow detailed instructions. Researchers have been exploring how to improve the data sets used to train LLMs to improve the performance and adaptability of the models.
Traditional evolution methods such as Evol-Instruct rely on evolution rules specified by human experts, which is not only expensive and time-consuming, but also requires redesign of the method when adapting to new tasks. In contrast, Auto Evol-Instruct realizes the automated evolution process by first using LLMs to analyze input instructions and independently design the initial method of evolution rules. Subsequently, the evolution method is iteratively optimized through optimizer LLMs to identify and solve problems during the evolution process to ensure the complexity and stability of the final evolution instructions.
Auto Evol-Instruct utilizes LLMs to design evolution methods by automatically analyzing input instructions and formulating evolution rules, thereby increasing the complexity and diversity of data sets.
In terms of performance evaluation, Auto Evol-Instruct performs well in multiple benchmark tests. For example, by fine-tuning Mixtral-8x7B using only 10K evolved ShareGPT data, the framework achieved 8.09 points on MT-Bench and 91.4 points on AlpacaEval, surpassing GPT-3.5-Turbo and WizardLM-70B, and competing with Claude2.0 is equivalent.
Furthermore, by using only 7K evolved GSM8K training data, the framework achieves 82.49 points on GSM8K. In terms of code generation, by fine-tuning DeepSeek-Coder-Base-33B using 20K evolved Code Alpaca, the framework achieves a score of 82.49 on HumanEval. It achieved a score of 77.4, surpassing other competing models.
It can be seen that this new framework performed well in multiple benchmark tests, including MT-Bench, AlpacaEval, GSM8K and HumanEval, demonstrating its potential in improving instruction following, mathematical reasoning and code generation capabilities.
Paper address: https://arxiv.org/abs/2406.00770
Highlights:
Auto Evol-Instruct is a fully automatic AI framework that can automatically analyze and evolve guidance data sets without human intervention.
The framework effectively increases the complexity and diversity of data sets by optimizing the evolution method, thereby enhancing the performance and adaptability of LLMs in various tasks.
The results of Auto Evol-Instruct demonstrate a method for guiding the evolution of data sets through automation.
The emergence of the Auto Evol-Instruct framework marks a major innovation in the evolution method of LLMs training data. Its automated and efficient features will greatly promote the development of LLMs and provide strong support for building more powerful and adaptable AI models. Relevant papers have been published, and interested readers can study them in depth.